Projecting DiffusionGemma distributions to the top item retains performance, countering fears that its diffusion steps create opaque serial depth. This finding strengthens research by Engels et al. regarding model transparency. It suggests that latent reasoning in diffusion-based text generation remains interpretable. Practitioners can monitor these internal states without significant loss in output quality.